feat(analyses): compute analyses in transformed scale space

[jkrumbiegel]

People often plot data on a log y axis and want a linear or smooth fit on top, but if you treat the data as gaussian in log space the fit has to be computed in log space too. Setting axis = (; yscale = log10) doesn't do that, it's a Makie axis attribute that warps the coordinates at draw time, after the analysis has already run in data space, so the fit is computed on the raw values and bends away from the data on the log axis.

This PR makes the scale set via scales actually affect analyses. scales(Y = (; scale = log10)) fits in the transformed space and back-transforms the result, so the fit lines up with the log-scaled display. It's a first-class scale, distinct from the display-only axis attribute.

That's the same split ggplot2 makes between scale_y_log10() (transforms before the stat) and coord_trans(y = "log10") (display only), so this is well-established practice. Quick check that AoG with scales(Y = (; scale = log10)) matches ggplot2's scale_y_log10() on identical data (left AoG, right ggplot2; lm top, loess bottom):

t = repeat(0:8, inner = 3)
conc = 100 .* 10 .^ (-0.11 .* t .+ 0.08 .* randn(length(t)))
spec = data((; t, conc)) * mapping(:t, :conc) * (visual(Scatter) + linear())

draw(spec, scales(Y = (; scale = log10)))

ggplot(df, aes(t, conc)) +
    geom_point() +
    geom_smooth(method = "lm") +
    scale_y_log10()

Covers linear, smooth, density, histogram, expectation and filled_contours. The transform sits inside the existing unit-stripping bracket so units still compose, and data outside the scale's domain (log of zero or a negative) errors early, naming the analysis and the offending value.

One consequence worth calling out: the positional aesthetics an analysis maps to (which arg is x, which is y) were already mostly assumed fixed in the existing code. This just turns that into a hard requirement, because the transform has to be chosen up front based on which aesthetic each positional lands on. So if a downstream visual changes the plot type or orientation such that data moves onto or off a scaled aesthetic the analysis didn't fit for, that's now an error instead of a silent mismatch.

This only kicks in when such a scale is actually set. Without one (or with identity), no pretransform happens, so there's no possible mismatch and nothing is checked: plots that worked before this PR behave exactly as before.

histogram also gains a direction keyword for 1D histograms, which sets the bar orientation and picks the aesthetic (so the scale space) the bins are computed in. Previously you'd flip orientation with a downstream visual(direction = :x), which can't carry the binning into the right scale space, so it lives on the analysis instead, like density's direction.